The Reserve Bank Is Learning to Read Your Business
Central banks now use AI to turn conversations with firms into inflation data, and South Africa’s liaison programme is just beginning.
Sign up for my Substack daily AI newsletter here.
Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/
When a central bank official sits across the table from a company executive and asks how prices are moving, the answer used to become a line in a report that a human read and filed. That is changing. Artificial intelligence can now read thousands of those conversations at once, score what firms say about their prices, and work out why those prices are rising. South Africa’s Reserve Bank began gathering exactly this kind of business intelligence only last year, which places it at the start of a shift that larger central banks are already well into. What companies tell the Bank is quietly becoming data, and that deserves every board’s attention.
CONTEXT AND BACKGROUND
The clearest demonstration comes from the Bank of England. Three of its staff used a large language model to read 66,820 reports from company visits written between 2005 and 2026, scoring the price changes firms described and splitting them into six drivers: energy, other input costs, labour, demand, profit margins and strategic pricing, meaning the matching of competitors. The resulting measure tracks consumer inflation closely and sometimes turned earlier than the Bank’s own staff assessments, as in 2016 and 2021.
South Africa is living through the kind of episode this method is built to diagnose. Headline inflation rose from 3.1% in March 2026 to 5% in June and sat at 4.4% in August, driven by an oil shock, and the Reserve Bank expects it to peak at 5.4% in the final quarter, well above the new 3% target that prompted a rate increase in September.
INSIGHT AND ANALYSIS
The Reserve Bank started its own business liaison programme with pilot interviews in August 2025, modelled on the Reserve Bank of Australia. Its contacts already report that strong competition and cheaper imports are limiting how much of their fuel and transport costs they can pass on, and that wage growth sits above target while total staff costs stay contained through productivity and attrition. The Bank is candid that the sample is not yet nationally representative, and it cross-checks claims by following up with other firms such as logistics operators (South African Reserve Bank, as above). A programme this young is a rare opportunity. It can be built from the start to be read well by machines, with consistent questions, a representative spread of firms and firm rules over confidential data, rather than retrofitted later.
The caution is that the machine is not a neutral reader. A separate Bank of England paper found that a language model asked to form inflation views lacked any consistent model of inflation, overreacted to visible items such as food and energy, and drifted upward as newer versions were released, which means every output needs validation.
I have previously written about the worldview that arrives with any AI system a buyer adopts, and the same warning applies here: the judgement must stay human, and accountability cannot be handed to a model. Elsewhere on the continent, the Bank of Ghana already runs an AI and big-data system it calls e-Inflation to nowcast prices, and uses text mining in its forecasting.
IMPLICATIONS
For boards, the first lesson is that corporate language is now economic data. What executives say about margins, wages and pricing power in earnings calls, investor briefings and liaison meetings is increasingly read and scored at scale, and it feeds into the assessments that set interest rates. The second is that firms can turn the same technique inward, reading their own supplier correspondence, branch reports and sales notes to see cost pressures forming before they reach the income statement.
Governance is where the value is won or lost. Sending confidential firm-level conversations to a foreign commercial model raises real questions of confidentiality and sovereignty, and tools trained mostly on advanced economies may misread South African firms. A Federal Reserve study found AI readings of central bank communication useful but dominated by rich-country material. The head of the Bank for International Settlements has argued that smaller economies should build applications tuned to their own data rather than import whole systems.
CLOSING TAKEAWAY
The advance here is not a better inflation forecast. It is the ability to diagnose why prices are moving, by turning thousands of ordinary business conversations into a consistent picture. South Africa has come to this late, which is its advantage, because it can build the foundations properly while the sample is still small. The task is to design the programme and its governance with care now, to keep skilled people in charge of what the models conclude, and to remember that behind every data point is a real firm describing a real pressure. Read well, that is a gift to policy. Read carelessly, it is a new way to be confidently wrong.
Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net
Comments